Research map / Label-efficient and robust learning / Recognition and long-tail
FreqDebias: Towards Generalizable Deepfake Detection via Consistency-Driven Frequency Debiasing
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Identifies spectral bias in deepfake detectors and proposes FreqDebias, combining frequency-domain Forgery Mixup augmentation with dual (CAM-based local and vMF hyperspherical global) consistency regularization to improve cross-domain generalization.
Problem
Deepfake detectors overfit to dominant frequency bands in training data and generalize poorly to unseen forgery types.
Approach
Fo-Mixup swaps amplitude spectra within clustered dominant frequency bands (selected via OHEM) to diversify frequency content while keeping phase/spatial structure, and dual consistency regularization enforces CAM consistency (local) and vMF hyperspherical distribution matching (global) between original and synthesized forgery domains.
Key results
- Cross-domain average AUC of 82.88% on FF++-trained detector across CDFv1, CDFv2, DFD, DFDCP, DFDC, outperforming LSDA (82.56%) and prior SOTA.
- Ablation: Fo-Mixup plus consistency losses raises CDFv2 AUC from 69.2 (baseline) to 83.6 and DFDCP from 61.3 to 82.4.
- Robustness on FF++ corruptions averages 97.6% video-level AUC, comparable to CADDM (97.4%).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add Fo-Mixup augmentation (ablation) | FreqDebias detector · Celeb-DF-v2 | AUC: 69.2 -> 74.8 | Helped |
| Add attention and hyperspherical consistency regularization alongside Fo-Mixup (ablation) | FreqDebias detector · Celeb-DF-v2 | AUC: 74.8 -> 83.6 | Helped |
| Add confidence sampling of synthesized samples (ablation) | FreqDebias detector · Celeb-DF-v2 | AUC: 81.7 -> 83.6 | Helped |
| Add attention consistency regularization (ablation) | FreqDebias with Fo-Mixup · Celeb-DF-v2 | AUC: 79.1 -> 83.6 | Helped |
| Add hyperspherical consistency regularization (ablation) | FreqDebias with Fo-Mixup · Celeb-DF-v2 | AUC: 80.4 -> 83.6 | Helped |
| Replace Fo-Mixup with amplitude mix (ablation) | FreqDebias · Celeb-DF-v2 | AUC: 83.6 -> 78.5 | Hurt |
| Replace Fo-Mixup with amplitude swap (ablation) | FreqDebias · Celeb-DF-v2 | AUC: 83.6 -> 76.0 | Hurt |
| Set confidence-sampling λ to 0.80 instead of 0.50 (ablation) | FreqDebias · Celeb-DF-v2 | AUC: 83.6 -> 82.1 | Hurt |
| Set confidence-sampling λ to 0.65 instead of 0.50 (ablation) | FreqDebias · Celeb-DF-v2 | AUC: 83.6 -> 82.9 | Hurt |
| Set confidence-sampling λ to 0.35 instead of 0.50 (ablation) | FreqDebias · Celeb-DF-v2 | AUC: 83.6 -> 82.4 | Hurt |
| Replace ResNet-34 with ResNet-50 (ablation) | FreqDebias · Celeb-DF-v2 | AUC: 83.6 -> 83.9 | Helped |
| Replace ResNet-34 with ConvNeXt (ablation) | FreqDebias · Celeb-DF-v2 | AUC: 83.6 -> 85.1 | Helped |
Surprising findings
- Standard detectors rely on very low-frequency bands for DeepFake forgeries but also on high-frequency bands for NeuralTextures, showing forgery-type-dependent spectral bias.
Limitations
- Relies on OHEM with a pre-trained forgery detector and K-means clustering hyperparameters (k, t, lambda) that are tuned on the evaluation setting.
- Training is limited to FF++ (HQ); generalization evaluated on the same benchmark families.
Experimental setup
Datasets: FaceForensics++ (HQ)
Benchmarks: FaceForensics++ Celeb-DF-v1 Celeb-DF-v2 DFD DFDCP DFDC FF++ cross-manipulation (DF, F2F, FS, NT) FF++ robustness corruptions
Metrics: frame-level AUC video-level AUC EER
Baselines: Xception Meso4 Capsule X-ray FFD F3Net SPSL SRM CORE RECCE SLADD IID
Builds on: DeepFakeBench OHEM CORE Frequency augmentation (amplitude mix/swap) DCL
Closest papers by meaning
- Frequency-Aware Deepfake Detection: Improving Generalizability through Frequency Space Domain Learning (AAAI 2024 · similarity 0.91)
- Exploring Unbiased Deepfake Detection via Token-Level Shuffling and Mixing (AAAI 2025 · similarity 0.90)
- FreqBlender: Enhancing DeepFake Detection by Blending Frequency Knowledge (NeurIPS 2024 · similarity 0.90)
- Transcending Forgery Specificity with Latent Space Augmentation for Generalizable Deepfake Detection (CVPR 2024 · similarity 0.90)
- Veritas: Generalizable Deepfake Detection via Pattern-Aware Reasoning (ICLR 2026 · similarity 0.89)
- Exposing the Deception: Uncovering More Forgery Clues for Deepfake Detection (AAAI 2024 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.