WildDFDet: Robust Unsupervised DeepFake Detection in the Wild

DeepFake is rapidly evolving driven by generative models, posing serious threats to digital media integrity and public trust. Although most supervised DeepFake detectors perform well under ideal conditions, their performance often declines remarkably in real-world forensic scenarios due to unlabeled, noisy-labeled, and limited training samples. A few unsupervised detectors avoid label dependence, but remain limited by the quality of pseudo-labels generated from handcrafted features and insufficient representation learning capability. To address these challenges, we propose WildDFDet, a robust unsupervised DeepFake detection framework for in-the-wild forensics. Firstly, we develop a Pseudo-Label Generator (PLG) based on a vision foundation model to generate pseudo-labels for wild data. Subsequently, we propose two collaborative mechanisms to improve robust representation learning under noisy supervision. Specifically, the first is Density-Aware Weighted Contrastive Learning (DAWCL), which promotes consistent natural face representation learning by assigning larger weights to samples located in high-density regions. The second is Density-Guided Label Correction (DGLC), which progressively refines noisy pseudo-labels through density-aware guidance. Extensive experiments demonstrate that WildDFDet achieves competitive performance across multiple benchmarks. Particularly, on our proposed in-the-wild benchmark, WildDFDet achieves an average AUC of 98.21%, outperforming existing methods by 27.68%–45.59%. The data is available at https://github.com/SHU-MMSecurity/WildFF