Focus on Finding Deepfakes: A Robust Proactive Detection Method Based on Orthogonal Moment Watermarking

Area of Science:

• Computer Science
• Digital Forensics
• Image Processing

Background:

• Deepfake technology poses significant challenges to digital media authenticity.
• Existing deepfake detection methods struggle with degraded image quality and conventional attacks.
• Robust and proactive detection mechanisms are crucial for combating sophisticated forgery techniques.

Purpose of the Study:

• To propose a novel watermarking-based proactive method for robust deepfake detection.
• To enhance feature extraction for more discriminative representations.
• To improve detection accuracy, especially under conventional attacks.

Main Methods:

• Embedding watermarks in the Fractional-order Quaternion Exponent Moments (FrQEMs) space for imperceptibility and robustness.
• Utilizing a Frequency Mamba (FreMamba) block to leverage frequency-domain correlations for enhanced feature extraction.
• Implementing a dual-branch framework with a watermark extractor and forgery discriminator, guided by knowledge distillation.

Main Results:

• The proposed method demonstrates superior deepfake detection accuracy on benchmark datasets.
• Watermark integrity is compromised specifically by deepfake attacks, remaining intact under conventional attacks.
• Outperforms state-of-the-art methods by over 5.3% in accuracy (ACC) when subjected to conventional attacks.

Conclusions:

• The watermarking-based proactive approach offers a robust solution for deepfake detection.
• The integration of FrQEMs watermarking and FreMamba feature extraction significantly enhances detection performance.
• The method provides reliable detection against deepfake attacks while maintaining robustness against conventional image manipulations.