HFGA-Net: High-frequency guided attention network for deepfake detection and localization

Area of Science:

• Computer Vision
• Digital Forensics
• Artificial Intelligence

Background:

• Deepfake technology poses significant trust and security risks due to sophisticated image and video manipulation.
• High-frequency components in forged images contain unique forensic artifacts crucial for detection.

Purpose of the Study:

• To develop a novel deepfake detection and localization method leveraging high-frequency information.
• To enhance the identification of forged faces by capturing fine-grained forgery cues.

Main Methods:

• Proposed the High-Frequency Guided Attention Network (HFGA-Net) for deepfake detection and localization.
• Introduced Frequency-Guided Patch-wise Attention (FGPA) to capture high-frequency information in shallow features.
• Developed Dynamic High-Frequency Multi-Stage Attention (DHFMA) for multi-stage feature analysis.
• Implemented a morphology-based mask generation method to address annotation scarcity.

Main Results:

• The HFGA-Net effectively leverages high-frequency information for superior forgery detection.
• FGPA and DHFMA modules successfully capture fine-grained forgery cues.
• The morphology-based mask generation method aids in training despite limited pixel-level annotations.
• Demonstrated strong generalization capabilities across various datasets for both detection and localization.

Conclusions:

• The proposed HFGA-Net offers a robust solution for deepfake detection and localization.
• The method's reliance on high-frequency artifacts provides a reliable approach to identifying manipulated media.
• The developed techniques show promise in combating the growing threat of deepfakes.