Research map / Label-efficient and robust learning / OOD and anomaly detection
FakeDiffer: Distributional Disparity Learning on Differentiated Reconstruction for Face Forgery Detection
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 FakeDiffer, a face forgery detector that learns real-fake distributional disparity via differentiated real/fake reconstruction with a shared encoder and information interaction fusion, to generalize to unseen manipulations.
Problem
Face forgery detectors overfit to known fake manipulation patterns and generalize poorly to unseen forgery types and datasets.
Approach
Redefines fake learning as real-fake distributional disparity learning: a unified encoder trained with two separate decoders (real and fake reconstruction) plus error-map encoders, whose outputs are fused by a learned information interaction module with the classifier.
Key results
- Intra-testing on FF++ C40 reaches 96.54% AUC, exceeding RECCE's 95.02%.
- Cross-dataset training on FF++ C40 gives 68.46% AUC on WildDeepfake, exceeding RECCE's 64.31% by about 4 points.
- Cross-manipulation testing on FF++ improves C-Avg to 72.00 for DF-trained models, above RECCE's 70.76.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use FakeDiffer instead of the strongest listed baseline | Xception-based face forgery detection · FF++ (c40) | AUC (%): 96.97 -> 96.54 | Hurt |
| use FakeDiffer instead of the strongest listed baseline | Xception-based face forgery detection · WildDeepfake | AUC (%): 93.27 -> 93.31 | Helped |
| use FakeDiffer instead of RECCE | face forgery detection trained on FF++ C40 · WildDeepfake | AUC: 64.31 -> 68.46 | Helped |
| use only Xception representation rather than the full FakeDiffer information sources (ablation) | FakeDiffer · FF++ C23; train F2F+FS+NT, test DF C23 | AUC: 98.47 -> 93.35 | Hurt |
| split reconstruction into real and fake categories rather than reconstructing without the split (ablation) | FakeDiffer · FF++ C23; train F2F+FS+NT, test DF C23 | AUC: 96.72 -> 98.47 | Helped |
| add real distributional information (ablation) | FakeDiffer · FF++ C23; train DF+FS+NT, test F2F C23 | AUC: 95.84 -> 98.36 | Helped |
| add fake distributional information (ablation) | FakeDiffer · FF++ C23; train DF+FS+NT, test F2F C23 | AUC: 94.46 -> 98.36 | Helped |
| use information interaction instead of adding real-inf and inva (ablation) | FakeDiffer · FF++ C23; train F2F+FS+NT, test DF C23 | AUC: 97.69 -> 98.47 | Helped |
| use information interaction instead of adding fake-inf and inva (ablation) | FakeDiffer · FF++ C23; train F2F+FS+NT, test DF C23 | AUC: 97.54 -> 98.47 | Helped |
| use information interaction instead of adding all three information sources (ablation) | FakeDiffer · FF++ C23; train F2F+FS+NT, test DF C23 | AUC: 97.50 -> 98.47 | Helped |
| include fake distributional information in information interaction (ablation) | FakeDiffer · FF++ C23; train DF+FS+NT, test F2F C23 | AUC: 97.82 -> 98.36 | Helped |
| include real distributional information in information interaction (ablation) | FakeDiffer · FF++ C23; train DF+FS+NT, test F2F C23 | AUC: 97.73 -> 98.36 | Helped |
Surprising findings
- Overfitting in face forgery detectors appears to mainly stem from known fake types rather than real images, as shown by the smaller performance drop when only fake types are unseen.
Limitations
- Gains over prior methods are modest on some intra-testing settings (e.g., CFM has higher AUC on FF++ C40 in the paper's own comparison, and Celeb-DF intra gains are small).
- Generalization evaluated only on FF++, Celeb-DF and WildDeepfake; no evaluation on diffusion-based or other newer forgery types.
Experimental setup
Datasets: FaceForensics++ Celeb-DF WildDeepfake
Benchmarks: FaceForensics++ (C23) FaceForensics++ (C40) Celeb-DF WildDeepfake
Metrics: Accuracy AUC EER
Baselines: Xception RFM Add-Net F3-Net MultiAtt RECCE Freq-SCL CFM DisGRL UniAttack
Builds on: RECCE Xception Multi-task manipulation detection
Closest papers by meaning
- DiffusionFake: Enhancing Generalization in Deepfake Detection via Guided Stable Diffusion (NeurIPS 2024 · similarity 0.89)
- FreqBlender: Enhancing DeepFake Detection by Blending Frequency Knowledge (NeurIPS 2024 · similarity 0.89)
- From Specificity to Generality: Revisiting Generalizable Artifacts in Detecting Face Deepfakes (NeurIPS 2025 · similarity 0.87)
- Critical Forgetting-Based Multi-Scale Disentanglement for Deepfake Detection (AAAI 2025 · similarity 0.87)
- DF40: Toward Next-Generation Deepfake Detection (NeurIPS 2024 · similarity 0.87)
- Exposing the Deception: Uncovering More Forgery Clues for Deepfake Detection (AAAI 2024 · similarity 0.87)
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.