Phase4DFD: Phase-Aware Deepfake Detection

• Phase4DFD is a deepfake detection framework that integrates explicit phase-magnitude modeling with multi-domain frequency analysis to reveal subtle artifacts.
• It enhances conventional RGB inputs by augmenting them with Fourier magnitude and local texture descriptors while utilizing a phase-aware attention module.
• Empirical results demonstrate that Phase4DFD outperforms spatial-only and magnitude-only methods on benchmark datasets with efficient computational overhead.

Phase4DFD is a deepfake detection framework that leverages multi-domain frequency analysis, integrating explicit phase-magnitude modeling with a learnable attention mechanism. It augments conventional RGB spatial inputs with Fourier magnitude and local texture descriptors, and employs a phase-aware attention module that targets frequency patterns most indicative of synthetic manipulation. This design is developed to address the limitations of spatial-only and magnitude-only detectors, achieving state-of-the-art performance with efficient computational overhead (Lin et al., 9 Jan 2026).

1. Motivation for Frequency-Domain and Phase Analysis

Recent advances in generative models, including GANs and diffusion networks, have diminished the efficacy of spatial-domain deepfake detectors relying on surface-level cues such as texture or geometry. These synthesis methods obscure spatial artifacts, making detection increasingly challenging. Frequency-domain representations expose latent manipulation cues, as generative pipelines introduce subtle irregularities in the Fourier spectrum. Prior deepfake detectors primarily exploit spectral magnitude; however, phase encodes structural alignment and content organization within an image. Authentic images typically display smoothly varying phase across adjacent frequencies, while generative synthesis disrupts these phase continuities. Explicit modeling of phase—alongside magnitude—enables the detection of nuanced artifacts inaccessible to magnitude-only approaches. Phase4DFD formulates a phase-aware input pipeline to guide feature extraction toward the most manipulation-sensitive frequency bands.

2. Construction of Multi-Domain Input Representation

Phase4DFD decomposes the standard RGB input $X\in\mathbb{R}^{3\times H\times W}$ into a five-channel augmented tensor $X^0\in\mathbb{R}^{5\times H\times W}$ by concatenating:

• Grayscale conversion: A single-channel intensity map $X_g\in\mathbb{R}^{1\times H\times W}$.
• FFT magnitude map:

$$M = \log\bigl|\mathrm{FFTShift}(\mathcal{F}(X_g))\bigr|, \qquad M\in\mathbb{R}^{1\times H\times W},$$

where $\mathcal{F}(\cdot)$ is the 2D Fourier Transform, FFTShift centralizes the DC component, and log-stabilization normalizes magnitude values.

• Differentiable LBP map: Local Binary Pattern descriptor $L\in\mathbb{R}^{1\times H\times W}$, sensitive to local texture transitions associated with synthetic manipulation.
• Channel concatenation:

$$X^0 = \text{concat}(X,\,M,\,L) \in \mathbb{R}^{5\times H\times W}.$$

This scheme synthesizes complementary spatial, spectral, and textural information, facilitating the learning of manipulation detectors robust to artifact suppression in any domain.

3. Phase-Aware Input Attention Mechanism

Phase4DFD integrates a novel input-level attention module exploiting phase-magnitude relationships. The normalized phase spectrum is computed: $$\Phi = \text{Norm}\Bigl(\angle\,\mathrm{FFTShift}(\mathcal{F}(X_g))\Bigr), \quad \Phi\in\mathbb{R}^{1\times H\times W},$$ where $\angle(\cdot)$ extracts phase and Norm scales to $[0,1]$.

Both $\Phi$ and $M$ are processed by parallel convolutional branches ($3\times3$ Conv → BN → ReLU), yielding feature tensors $F_\Phi$ and $F_M$. These are concatenated, projected via $1\times1$ convolution, and squashed by a sigmoid activation to produce the attention tensor: $A^0\in\mathbb{R}^{5\times H\times W}.$ Elementwise modulation produces the attended augmented input: $\widetilde{X}^0 = X^0 \odot A^0.$

At the frequency-bin level $(i, j)$, attention weights are given by: $$\alpha_{i,j} = \frac{\exp(f(\Phi_{i,j},\,M_{i,j}))}{\sum_{p,q}\exp(f(\Phi_{p,q},\,M_{p,q}))},$$ where $f$ is a small neural fusion module. High attention values are assigned to bins exhibiting abnormal phase-magnitude pairing, as is typical of generative artifacts. This directs feature extraction toward spectral regions with the highest likelihood of manipulation.

4. Backbone Network and Feature Refinement

The attended input $\widetilde{X}^0$ ($5$ channels) passes through a $1\times1$ channel adapter, reducing it to the conventional three-channel format ($X^a\in\mathbb{R}^{3\times H\times W}$). The encoder architecture is BNext-M, a compact hierarchical convolutional network that expands receptive fields efficiently.

An optional feature-level channel–spatial attention module (CBAM style) further processes the output features $F\in\mathbb{R}^{2048\times 7\times 7}$ via:

• Channel attention:

$$A_c = \sigma\left(\mathrm{MLP}(\mathrm{GAP}(F))\right)\in\mathbb{R}^{2048\times 1\times 1}$$

• Spatial attention:

$$A_s = \sigma\left(\mathrm{Conv}_{7\times7}([\mathrm{AvgPool}(F);\,\mathrm{MaxPool}(F)])\right)\in\mathbb{R}^{1\times 7\times 7}$$

• Feature refinement:

$F_s = (F\odot A_c)\odot A_s$

Empirical evaluation reveals that core input-level phase-aware attention provides the dominant performance improvements, with feature-level attention offering only marginal gains.

5. Training Protocol and Datasets

Phase4DFD is evaluated on two benchmark datasets:

Dataset	Image Count	Real / Fake Distribution	Resolution	Partitioning
CIFAKE	120,000	60K real, 60K Stable Diff.	32×32 → 224×224	100K train / 20K test
DFFD	≈300,000	≈58K real, ≈240K PGGAN/StyleGAN	192×192	50% train / 5% val / 45% test

• Augmentation: Random flip, rotation ($\pm15^\circ$), color jitter, resized crop—performed prior to FFT/LBP extraction for domain consistency.
• Normalization: Standard ImageNet normalization after channel adaptation.
• Optimization: AdamW, cosine-annealed learning rate.
• Loss function: Weighted blend of BCE and Focal Loss:

$$\mathcal{L}_{\rm train} = 0.7 \mathcal{L}_{\rm BCE} + 0.3 \mathcal{L}_{\rm Focal}$$

where $$\mathcal{L}_{\rm BCE} = -[w_{\rm pos} y\log p + (1-y)\log(1-p)]$$ with $w_{\rm pos}=N_{\rm real}/N_{\rm fake}$, $$\mathcal{L}_{\rm Focal} = -\alpha (1-p)^\gamma y\log p$$, $\gamma=2$.

• Training schedule: Two-stage strategy—initially freezing BNext-M for $5$ (CIFAKE) or $10$ (DFFD) epochs, optimizing only attention and classifier (lr=$1\times10^{-3}$), followed by fine-tuning all modules for $15$ epochs (backbone lr=$1\times10^{-4}$, others $1\times10^{-3}$).

6. Experimental Performance and Ablation Studies

Phase4DFD achieves superior accuracy and AUC metrics compared to Xception, VGG16, and baseline BNext-M detectors:

Model	DFFD Accuracy	DFFD AUC	CIFAKE Accuracy	CIFAKE AUC
BNext-M (baseline)	98.75%	99.92	97.35%	99.62
Phase4DFD	99.46%	99.95	98.62%	99.88

On CIFAKE, F1-scores are balanced (98.62) across both real and fake classes, reflecting robust discriminative power.

Ablation studies on DFFD reveal:

• RGB-only: 99.23% accuracy.
• Adding FFT magnitude: +0.03%; adding LBP: +0.01%. Joint addition without phase attention degrades performance.
• Feature-level attention (CBAM): accuracy lifts to 99.18%.
• Input-level phase-aware attention: accuracy rises to 99.46%, substantiating the complementary, non-redundant utility of explicit phase-magnitude modeling at the input stage.

This suggests that revisiting fundamental signal properties—such as phase continuity—can meaningfully enhance manipulation detection without increasing model complexity.

7. Implications and Future Prospects

Phase4DFD demonstrates that phase-aware, multi-domain attention architectures can substantially outperform traditional spatial and magnitude-based deepfake detectors without incurring significant computational cost. A plausible implication is that future research on image forensics and synthetic media authentication will increasingly emphasize joint frequency-phase representations and input-level attention mechanisms. The empirical evidence supporting the non-redundancy of explicit phase modeling advocates for systematic inclusion of phase analysis in frequency-domain learning pipelines. Further exploration could probe the generalization of this approach to non-facial domains, adversarial robustness, and real-time applications.