An Efficient Active Learning based Method for Deepfake Vidoes Model Attribution
Xiaotian Si, Weiqiang Jiang, Linghui Li, Xiaoyong Li, Kaiguo Yuan, Zhongyuan Guo · 2023
To enhance the forensic investigation of deepfake face-swap videos, it is essential to attribute the specific generation model used to create these videos. Despite the remarkable progress made in data-driven approaches, recent algorithms continue to encounter challenges in learning from limited annotated data, thereby limiting their performance. In this paper, we present a novel active learning framework for the model attribution of deepfake videos. Specifically, our approach leverages active learning to select the most informative videos, thereby effectively reducing the annotation effort. To achieve this, we propose a query function based on clustering, where the clustering features are generated by a projection network trained through supervised contrastive learning. Extensive experiments demonstrate that our method achieves state-of-the-art performance in model attribution of deepfake videos, surpassing other baseline methods in the field of active learning.