Research map / Label-efficient and robust learning / Recognition and long-tail
DFD-HR: Generalizable Deepfake Detection via Hierarchical Routing Learning
Jiamu Sun, Zhiyuan Yan, Ke-Yue Zhang, Taiping Yao, Shouhong Ding
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add Early Layer Pruning (ablation) | DFD-HR (CLIP ViT-L/14) · CDF-v2/FaceShifter avg | video AUC (%): 90.3 -> 91.3 | Improved |
| add Token Selection with Spearman rank loss (ablation) | DFD-HR · CDF-v2/FaceShifter avg | video AUC (%): 91.3 -> 92.4 | Improved |
| add token-level Expert Routing (ablation) | DFD-HR · CDF-v2/FaceShifter avg | video AUC (%): 92.4 -> 92.9 | Improved |
| add Multi-Scale Fusion (ablation) | DFD-HR · CDF-v2/FaceShifter avg | video AUC (%): 92.9 -> 93.6 | Improved |
| replace Token Selection with Sparse Attention (ablation) | DFD-HR · CDF-v2/FaceShifter avg | video AUC (%): 91.3 -> 91.0 (vs ELP baseline) | Hurt |
| replace Token Selection with PCA or FastV (ablation) | DFD-HR · CDF-v2/FaceShifter avg | video AUC (%): 91.7 and 91.8 vs 92.4 for ours | Mixed |
| replace Multi-Scale Fusion with RINE or Multi CLS Pooling or D3 (ablation) | DFD-HR · CDF-v2/FaceShifter avg | video AUC (%): baseline 90.3 -> 84.2 (RINE), 89.0, 89.1 | Hurt |
| remove Gumbel-Sigmoid hard routing in Early Layer Pruning (ablation) | DFD-HR · CDF-v2/FaceShifter avg | video AUC (%): 93.6 -> 74.8 | Hurt |
| share layer judge across layers (ablation) | DFD-HR · CDF-v2/FaceShifter avg | video AUC (%): 93.6 -> 91.8 | Hurt |
| use noisy gate instead of Top2 routing or balance loss in experts (ablation) | DFD-HR · CDF-v2/FaceShifter avg | video AUC (%): 92.9 (Top2), 92.4 (Balance) -> 93.6 | Improved |
| increase number of experts beyond 4 (ablation) | DFD-HR · CDF-v2/FaceShifter avg | video AUC (%): 93.6 (4) -> 93.1 (6), 93.0 (8) | Hurt |
| increase rank loss weight lambda1 from 0.1 to 1.0 (ablation) | DFD-HR · CDF-v2/FaceShifter avg | video AUC (%): 93.6 -> 92.5 | Hurt |
Closest papers by meaning
- Veritas: Generalizable Deepfake Detection via Pattern-Aware Reasoning (ICLR 2026 · similarity 0.91)
- Towards More General Video-based Deepfake Detection through Facial Component Guided Adaptation for Foundation Model (CVPR 2025 · similarity 0.91)
- Transcending Forgery Specificity with Latent Space Augmentation for Generalizable Deepfake Detection (CVPR 2024 · similarity 0.91)
- Standing on the Shoulders of Giants: Reprogramming Visual-Language Model for General Deepfake Detection (AAAI 2025 · similarity 0.90)
- D^3: Scaling Up Deepfake Detection by Learning from Discrepancy (CVPR 2025 · similarity 0.90)
- FreqDebias: Towards Generalizable Deepfake Detection via Consistency-Driven Frequency Debiasing (CVPR 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.
