Balancing the Unbalanced: Improving Fairness in Deepfake Detection with Imbalanced Datasets
Year: 2025
Type: conference-paper
Abstract: The rapid progress of AIGC has led to the widespread availability of highly realistic synthetic face images, raising serious concerns about identity misuse and public trust. As a countermeasure, deepfake detection has emerged as a critical research field. However, recent studies reveal that existing detection models exhibit significant performance disparities across different races. These biases are primarily attributed to imbalanced training datasets, where Caucasian faces are overrepresented. This dual bias—at both the data and generation quality—undermines the fairness and generalisability of deepfake detection systems. To mitigate demographic bias in deepfake detection, we propose DFAD, a plug-and-play debiasing module designed to decouple demographic attributes from task-relevant representations. DFAD integrates dynamic focal loss reweighting and adversarial debiasing to encourage fairer feature learning without requiring explicit demographic annotations. We incorporate DFAD into two state-of-the-art detection models, CADDM and DFGaze, and evaluate performance on the FF++ dataset. Experimental results demonstrate that DFAD consistently improves fairness across racial groups, while maintaining competitive detection accuracy.
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
SDG: Reduced inequalities
Open Access status: hybrid
OpenAlex ID: W7127051097