Black-Box Adaptation for Deepfake Detection via Local Relation Guided AUC Optimization

Deepfake technologies pose a growing threat to the Internet of Things (IoT), enabling identity spoofing and the spread of misinformation. Although numerous face forgery detectors have been developed to counter these risks, their real-world deployment is often limited by inherent dataset biases. Existing domain adaptation techniques offer potential remedies, but typically rely on access to raw source data and employ data-dependent alignment strategies under a transductive learning paradigm, raising substantial privacy concerns for source domain individuals. This study revisits the problem from the perspective of black-box domain adaptation (BDA) and introduces a detection framework that leverages only the predictions from the source model. The method is grounded in a local relation-guided area under the ROC curve (AUC) optimization strategy, which leverages the robustness of AUC-based objectives in noisy environments while addressing the limitations of conventional AUC optimization, particularly its vulnerability to confirmation bias and reliance on a fixed decision threshold. To this end, two key components are introduced. First, a nearest-neighbor calibration mechanism is presented, where the local relation feature (LRF), a parameter-free representation, captures differences between real and fake images without favoring specific forgery types, thereby reducing bias inherited from the source model. Second, a Gaussian mixture model (GMM)-based adaptive thresholding scheme is employed to address asymmetric predictions by dynamically determining the optimal decision boundary for real and fake images. Experiments across multiple datasets show that our approach achieves superior generalization compared to state-of-the-art (SOTA) methods. Moreover, the framework is compatible with a broad range of source and target model configurations to enhance detection performance.