New lightweight fusion method boosts deepfake detection by 4.4% with just 292 extra parameters
Researchers show two handcrafted cues outperform massive models for spotting fake videos...
Deep Dive
Researchers developed a lightweight fusion method for video face forgery detection. Adding only 292 parameters to the standard Xception backbone (21.9M total), their approach combines low-frequency wavelet-denoised features with phase-spectrum or local binary patterns. On FaceForensics++ and DFDC-Preview, AUC improved by 3.8% and 4.4% respectively—beating larger models like F3Net (22.5M) and SRM (55.3M) across eight benchmarks without extra data or augmentation.
Key Points
- Method adds only 292 parameters to standard Xception backbone (21.9M total), far smaller than competitors like SRM (55.3M).
- AUC improves by 3.8% on FaceForensics++ (to 78.6%) and 4.4% on DFDC-Preview (to 74.9%).
- Outperforms F3Net, SRM, and SPSL in eight benchmarks without extra data or test-time augmentation.
Why It Matters
Efficient, high-accuracy deepfake detection for real-time video verification, challenging the need for massive models.