FreqMamba: Spatial-Frequency Fusion and State Space Sequence Modeling for Deepfake Detection.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Digital Forensics

Background:

• Deepfake technology poses significant threats to social trust and personal privacy.
• Current deepfake detection methods struggle with cross-domain generalization due to limitations in spatial and frequency domain analyses.
• Imperceptible forgery artifacts and ineffective feature integration hinder existing detection algorithms.

Purpose of the Study:

• To develop an advanced face forgery detection framework with strong cross-domain generalization capabilities.
• To address the limitations of existing spatial-domain and frequency-aware deepfake detection methods.
• To propose FreqMamba, an end-to-end framework integrating spatial semantic and frequency-domain features.

Main Methods:

• Proposed FreqMamba framework utilizing a CNN branch for spatial semantic features.
• Incorporated a hierarchical discrete wavelet transform (DWT) branch for fine-grained frequency artifact analysis.
• Employed a bidirectional vision state space model (Vim) for global sequence modeling with linear complexity and a gated late-fusion mechanism.

Main Results:

• FreqMamba achieved 0.7767 AUC on Celeb-DF v2, outperforming a CNN baseline by 5.05%.
• On the WildDeepfake dataset, FreqMamba attained 0.6993 AUC, exceeding a lightweight CNN baseline by 7.21%.
• Ablation studies and Grad-CAM visualizations confirmed the synergistic effects of DWT and Mamba, and the model's focus on critical forgery regions.

Conclusions:

• FreqMamba offers an effective solution for generalized face forgery detection across different domains.
• The integration of spatial semantics, frequency artifacts, and global sequence modeling enhances detection performance.
• The proposed framework significantly improves cross-domain generalization, crucial for real-world deepfake detection applications.