Generalizable Deepfake Detection With Phase-Based Motion Analysis.

Area of Science:

• Computer Vision
• Digital Forensics
• Machine Learning

Background:

• DeepFake (DF) video detection methods leveraging temporal information outperform per-frame techniques.
• Existing temporal methods struggle with cross-dataset generalization and robustness to distortions.
• Issues include inaccurate motion estimation, landmark tracking, and susceptibility to adversarial attacks.

Purpose of the Study:

• To introduce PhaseForensics, a novel DeepFake video detection method.
• To enhance robustness against common distortions and adversarial attacks.
• To improve cross-dataset generalization capabilities in DeepFake detection.

Main Methods:

• Utilizing a phase-based motion representation of facial temporal dynamics.
• Leveraging temporal phase variations in band-pass frequency components of face regions.
• Employing band-pass filters for robust temporal dynamics estimation and defense against adversarial perturbations.

Main Results:

• PhaseForensics demonstrates improved distortion and adversarial robustness.
• Achieved state-of-the-art cross-dataset generalization.
• Reached 92.4% video-level AUC on the CelebDFv2 benchmark, outperforming prior methods (e.g., FTCN at 86.9%).

Conclusions:

• Phase-based motion analysis offers a robust approach for DeepFake video detection.
• The proposed method effectively addresses limitations of existing temporal detection techniques.
• PhaseForensics represents a significant advancement in reliable and generalizable DeepFake detection.