Research map / Label-efficient and robust learning / Recognition and long-tail
FreqBlender: Enhancing DeepFake Detection by Blending Frequency Knowledge
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.
Proposes FreqBlender, which blends structural frequency components (parsed by a learned Frequency Parsing Network over DCT maps) from fake faces into real faces to create frequency-aware pseudo-fakes that complement spatial blending (SBI) for DeepFake detection.
Problem
Existing pseudo-fake face generation for DeepFake detection blends faces only in the spatial domain and ignores frequency distribution, so learned detectors miss generic frequency-based forgery traces and generalize poorly to unseen forgeries.
Approach
A Frequency Parsing Network (shared encoder + three decoders) predicts soft DCT-domain masks for semantic, structural, and noise components, trained without ground truth via four auxiliary objectives (facial fidelity, authenticity-determinative, quality-agnostic, prior-integrity); structural components of fake faces are then blended into real faces, and this is applied on top of SBI-generated pseu
Key results
- Trained only on FF++, FreqBlender reaches 94.59 AUC on CDF, 87.56 on DFDCP and 86.14 on FFIW, improving over SBI by about 1.4 points on CDF and DFDCP.
- On the FF++ cross-manipulation test, FreqBlender averages 96.13 AUC versus 92.51 for SBI-raw.
- Adding FreqBlender to Face X-ray raises its average over the four tested sets from 69.73 to 72.81.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add frequency blending to SBI-based spatial pseudo-fakes | EfficientNet-b4 DeepFake detector · CDF | AUC (%): 93.18 -> 94.59 | Helped |
| add frequency blending to SBI-based spatial pseudo-fakes | EfficientNet-b4 DeepFake detector · DFDC | AUC (%): 72.42 -> 74.59 | Hurt |
| add frequency blending to SBI-based spatial pseudo-fakes | EfficientNet-b4 DeepFake detector · DFDCP | AUC (%): 86.15 -> 87.56 | Helped |
| add frequency blending to SBI-based spatial pseudo-fakes | EfficientNet-b4 DeepFake detector · FFIW | AUC (%): 84.83 -> 86.14 | Helped |
| add facial fidelity loss L_ff (ablation) | FreqBlender · CDF, DFDC, DFDCP Avg | AUC (%): 85.06 -> 85.58 | Helped |
| add authenticity-determinative loss L_ad (ablation) | FreqBlender · CDF, DFDC, DFDCP Avg | AUC (%): 84.20 -> 85.58 | Helped |
| add quality-agnostic loss L_qa (ablation) | FreqBlender · CDF, DFDC, DFDCP Avg | AUC (%): 85.32 -> 85.58 | Helped |
| add prior and integrity loss L_pi (ablation) | FreqBlender · CDF, DFDC, DFDCP Avg | AUC (%): 84.60 -> 85.58 | Helped |
| add FreqBlender to DSP-FWA (ablation) | DSP-FWA pseudo-fake generation · FF++, CDF, DFDCP, FFIW Avg | AUC (%): +0.12% | Helped |
| add FreqBlender to I2G (ablation) | I2G pseudo-fake generation · FF++, CDF, DFDCP, FFIW Avg | AUC (%): +0.84% | Helped |
| add FreqBlender to Face X-ray (ablation) | Face X-ray pseudo-fake generation · FF++, CDF, DFDCP, FFIW Avg | AUC (%): +3.08% | Helped |
| replace SBI with FreqBlender pseudo-fakes (ablation) | VGG16 detector · CDF, FF++, DFDCP, FFIW Avg | AUC (%): 83.16 -> 83.15 | Hurt |
Surprising findings
- Using only FreqBlender without spatial-blending pseudo-fakes (alpha=1) performs very poorly (e.g., 56.34 AUC on CDF), indicating that frequency blending alone lacks color-space knowledge.
- Under FGSM attack the method retains only 22.31 AUC, and BIM, PGD and CW reduce all methods to 0 AUC.
Limitations
- Assumes faces are forged by face-swapping techniques, so generalization to whole-face synthesis and attribute editing is untested.
- Gains on some networks are marginal or negative (e.g., VGG16 drops 0.01 points) and the method is sensitive to the alpha mixing probability.
- Vulnerable to white-box evasion attacks, with near-zero AUC under BIM, PGD and CW.
Experimental setup
Datasets: FaceForensics++ Celeb-DF DFDC DFDCP FFIW-10k
Benchmarks: FaceForensics++ Celeb-DF DFDC DFDCP FFIW-10k StyleGAN StyleGAN2
Metrics: AUC
Baselines: Two-branch DAM LipForensics FTCN SST DSP-FWA Face X-ray F3-Net LRL FRDM PCL+I2G DCL
Builds on: SBI FWA Face X-ray EfficientNet
Closest papers by meaning
- FreqDebias: Towards Generalizable Deepfake Detection via Consistency-Driven Frequency Debiasing (CVPR 2025 · similarity 0.90)
- Frequency-Aware Deepfake Detection: Improving Generalizability through Frequency Space Domain Learning (AAAI 2024 · similarity 0.90)
- DF40: Toward Next-Generation Deepfake Detection (NeurIPS 2024 · similarity 0.89)
- DiffusionFake: Enhancing Generalization in Deepfake Detection via Guided Stable Diffusion (NeurIPS 2024 · similarity 0.89)
- Exposing the Deception: Uncovering More Forgery Clues for Deepfake Detection (AAAI 2024 · similarity 0.89)
- FakeDiffer: Distributional Disparity Learning on Differentiated Reconstruction for Face Forgery Detection (AAAI 2025 · similarity 0.89)
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.