[Paper] Phase4DFD: Multi-Domain Phase-Aware Attention for Deepfake Detection

Source: arXiv - 2601.05861v1

Overview

The paper Phase4DFD introduces a new deepfake detection framework that goes beyond the usual pixel‑level analysis and taps into the frequency domain—specifically, the often‑ignored phase component of the Fourier transform. By marrying RGB images with magnitude, local binary pattern (LBP), and a learnable phase‑aware attention module, the authors achieve state‑of‑the‑art detection accuracy while keeping the model lightweight enough for real‑time deployment.

Key Contributions

• Phase‑aware attention: A novel input‑level module that highlights phase discontinuities—common by‑products of synthetic video generation—and guides the backbone toward the most telling frequency cues.
• Multi‑domain input fusion: Simultaneous feeding of RGB, FFT magnitude, and LBP maps, exposing manipulation artifacts invisible to spatial‑only methods.
• Efficient backbone: Integration with the BNext‑M architecture (and optional channel‑spatial attention) that delivers high accuracy with modest compute and memory footprints.
• Comprehensive evaluation: Superior performance on two large‑scale benchmarks (CIFAKE and DFFD) compared with both spatial and frequency‑only detectors.
• Ablation insights: Demonstrates that phase information contributes complementary, non‑redundant signals beyond magnitude‑only representations.

Methodology

1. Pre‑processing:
   • Input video frames are converted to three parallel representations:
     • RGB (standard color image).
     • FFT magnitude obtained via a Fast Fourier Transform, capturing the strength of each frequency component.
     • Local Binary Pattern (LBP) maps that encode fine‑grained texture cues.
2. Phase‑aware Attention Module:
   • The FFT also yields a phase map (the angle of each frequency component).
   • The module learns an attention mask that emphasizes regions where the phase shows abrupt changes—these are typical of generative artifacts such as stitching or interpolation.
   • The mask is applied before any deep feature extraction, effectively “pre‑filtering” the multi‑domain inputs.
3. Backbone Feature Extraction:
   • The attended multi‑domain tensor is fed into BNext‑M, a compact convolutional network designed for speed.
   • An optional channel‑spatial attention (CSA) block refines the semantic features by re‑weighting channel and spatial dimensions.
4. Classification Head:
   • A lightweight fully‑connected layer predicts the binary label (real vs. deepfake).
5. Training:
   • Standard cross‑entropy loss with data augmentation (random cropping, horizontal flip) and frequency‑domain augmentations (phase jitter) to improve robustness.

Results & Findings

• Accuracy boost: Adding phase‑aware attention yields ~2–3 % AUC improvement over magnitude‑only baselines.
• Efficiency: The full pipeline runs at ~45 FPS on a single RTX 3080, with <120 MB GPU memory—well within the limits for edge or streaming scenarios.
• Ablation: Removing the phase module drops performance to the level of magnitude‑only models, confirming that phase contributes unique information.
• Robustness: The model maintains high detection rates under common post‑processing (compression, resizing), indicating that phase cues survive typical distribution shifts.

Practical Implications

• Real‑time moderation: The low latency and modest hardware requirements make Phase4DFD suitable for live video platforms (e.g., streaming services, video conferencing) that need on‑the‑fly deepfake screening.
• Forensic tooling: Investigators can integrate the multi‑domain preprocessing pipeline into existing forensic suites to uncover subtle manipulations that evade visual inspection.
• Edge deployment: Because the backbone is lightweight, the approach can be packaged for mobile or embedded devices (e.g., smart cameras) to perform on‑device authenticity checks without sending raw footage to the cloud.
• Model‑agnostic augmentation: The phase‑aware attention module can be grafted onto other detection backbones (ResNet, EfficientNet), offering a plug‑and‑play upgrade path for teams already invested in different architectures.

Limitations & Future Work

• Phase sensitivity to extreme compression: While robust to moderate codecs, very low‑bitrate streams can distort phase information, slightly degrading detection.
• Generalization to unseen generation methods: The study focuses on two benchmark datasets; newer generative models (e.g., diffusion‑based video synthesis) may exhibit different phase signatures, requiring further validation.
• Explainability: Although the attention maps highlight phase discontinuities, a deeper interpretability analysis (e.g., linking specific artifacts to generation pipelines) is left for future research.
• Multi‑modal extensions: Incorporating audio or temporal consistency cues alongside phase‑aware frequency analysis could further harden detectors against sophisticated attacks.

Authors

• Zhen‑Xin Lin
• Shang‑Kuan Chen

Paper Information

• arXiv ID: 2601.05861v1
• Categories: cs.CV
• Published: January 9, 2026
• PDF: Download PDF