• Proposes a Full-Defense Framework (FDF) that integrates passive detection and proactive defense against deepfake.
• Introduces separable watermarks (SepMark) with a robust decoder for source tracing and a semi-robust decoder sensitive to malicious distortions.
• Employs cross-domain feature fusion of spatial and frequency channels to improve discrimination between deepfake content and watermark removal attacks.
• Achieves dual functionality: copyright protection and deepfake detection even when watermarks are absent, offering a comprehensive defense solution.