Exposing the Deception: Uncovering More Forgery Clues for Deepfake Detection

Exposing the Deception: Uncovering More Forgery Clues for Deepfake Detection

4 Mar 2024 | Zhongjie Ba1,2, Qingyu Liu1,2, Zhenguang Liu1,2 *, Shuang Wu3, Feng Lin1,2, Li Lu1,2, Kui Ren1,2
This paper addresses the limitations of current deepfake detection methods, which often overfit to specific regions and lack theoretical guarantees for extracting sufficient and relevant forgery clues. The authors propose a novel framework that captures broader forgery clues by extracting multiple non-overlapping local representations and fusing them into a global semantic-rich feature. The framework is based on information bottleneck theory, which ensures the orthogonality of local representations while preserving comprehensive task-relevant information. Additionally, a Global Information Loss is introduced to eliminate task-irrelevant information. Empirical results on five benchmark datasets demonstrate that the proposed method achieves state-of-the-art performance, outperforming existing methods in both in-dataset and cross-dataset settings. The method's effectiveness is validated through extensive experiments, ablation studies, and visualizations, highlighting its ability to uncover more forgery clues and improve generalizability.This paper addresses the limitations of current deepfake detection methods, which often overfit to specific regions and lack theoretical guarantees for extracting sufficient and relevant forgery clues. The authors propose a novel framework that captures broader forgery clues by extracting multiple non-overlapping local representations and fusing them into a global semantic-rich feature. The framework is based on information bottleneck theory, which ensures the orthogonality of local representations while preserving comprehensive task-relevant information. Additionally, a Global Information Loss is introduced to eliminate task-irrelevant information. Empirical results on five benchmark datasets demonstrate that the proposed method achieves state-of-the-art performance, outperforming existing methods in both in-dataset and cross-dataset settings. The method's effectiveness is validated through extensive experiments, ablation studies, and visualizations, highlighting its ability to uncover more forgery clues and improve generalizability.
Reach us at info@study.space
Understanding Exposing the Deception%3A Uncovering More Forgery Clues for Deepfake Detection